Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2018 Oct 30;14(10):e1006532.
doi: 10.1371/journal.pcbi.1006532. eCollection 2018 Oct.

Predicting Bioprocess Targets of Chemical Compounds Through Integration of Chemical-Genetic and Genetic Interactions

Free PMC article

Predicting Bioprocess Targets of Chemical Compounds Through Integration of Chemical-Genetic and Genetic Interactions

Scott W Simpkins et al. PLoS Comput Biol. .
Free PMC article


Chemical-genetic interactions-observed when the treatment of mutant cells with chemical compounds reveals unexpected phenotypes-contain rich functional information linking compounds to their cellular modes of action. To systematically identify these interactions, an array of mutants is challenged with a compound and monitored for fitness defects, generating a chemical-genetic interaction profile that provides a quantitative, unbiased description of the cellular function(s) perturbed by the compound. Genetic interactions, obtained from genome-wide double-mutant screens, provide a key for interpreting the functional information contained in chemical-genetic interaction profiles. Despite the utility of this approach, integrative analyses of genetic and chemical-genetic interaction networks have not been systematically evaluated. We developed a method, called CG-TARGET (Chemical Genetic Translation via A Reference Genetic nETwork), that integrates large-scale chemical-genetic interaction screening data with a genetic interaction network to predict the biological processes perturbed by compounds. In a recent publication, we applied CG-TARGET to a screen of nearly 14,000 chemical compounds in Saccharomyces cerevisiae, integrating this dataset with the global S. cerevisiae genetic interaction network to prioritize over 1500 compounds with high-confidence biological process predictions for further study. We present here a formal description and rigorous benchmarking of the CG-TARGET method, showing that, compared to alternative enrichment-based approaches, it achieves similar or better accuracy while substantially improving the ability to control the false discovery rate of biological process predictions. Additional investigation of the compatibility of chemical-genetic and genetic interaction profiles revealed that one-third of observed chemical-genetic interactions contributed to the highest-confidence biological process predictions and that negative chemical-genetic interactions overwhelmingly formed the basis of these predictions. We also present experimental validations of CG-TARGET-predicted tubulin polymerization and cell cycle progression inhibitors. Our approach successfully demonstrates the use of genetic interaction networks in the high-throughput functional annotation of compounds to biological processes.

Conflict of interest statement

I have read the journal's policy and the authors of this manuscript have the following potential competing interests: Authors Simpkins, Nelson, and Myers have licensed the CG-TARGET software for commercial use (


Fig 1
Fig 1. Overview of the integration of chemical-genetic and genetic interaction networks for bioprocess target prediction using CG-TARGET.
Chemical-genetic interaction profiles, obtained by measuring the sensitivity or resistance of a library of gene mutants to a chemical compound, are compared against genetic interaction profiles consisting of double mutant interaction scores. The resulting similarities are aggregated at the level of biological processes to predict the bioprocess(es) perturbed by the compound. Better agreement between chemical-genetic and genetic interaction profiles leads to stronger bioprocess predictions. Each blue box represents a negative chemical-genetic (i.e. sensitivity) or genetic interaction, while each black box represents the absence of an interaction. Stronger bioprocess predictions are depicted with a darker red.
Fig 2
Fig 2. Rate of compound discovery and control of the false discovery rate for the prediction of bioprocesses from chemical-genetic interaction profiles.
Perturbed bioprocesses were predicted using CG-TARGET for compounds, experimental controls (DMSO), and resampled chemical-genetic interaction profiles from the RIKEN and NCI/NIH/GSK datasets. (A) The number of compounds, experimental controls, and randomly resampled chemical-genetic interaction profiles discovered with at least one bioprocess prediction passing the given significance thresholds, for the RIKEN dataset. (B) DMSO and resampled profile-derived estimates of the false discovery rate of biological process predictions, for the RIKEN dataset, given the number of discovered compounds. Values were calculated from (A). (C-D) Same as (A-B), respectively, but for the NCI/NIH/GSK dataset.
Fig 3
Fig 3. Comparison of CG-TARGET performance versus gene-target enrichment.
Perturbed bioprocesses were predicted using both CG-TARGET and a method that calculated enrichment on the set of each compound’s 20 most similar genetic interaction profiles (“top 20”). (A) Bioprocess prediction false discovery rate estimates derived from resampled chemical-genetic interaction profiles, performed on compounds from the RIKEN dataset. (B) Precision-recall analysis of the ability to recapitulate gold-standard annotations within the set of top bioprocess predictions for ~4500 simulated compounds. Each simulated compound was designed to target one query gene in the genetic interaction network and thus inherited gold-standard biological process annotations from its target gene. (C) For each of 35 well-characterized compounds in the RIKEN dataset with literature-derived, gold-standard biological process annotations, we determined the rank of its gold-standard bioprocess within its list of predictions. The number of compounds for which a given rank (or better) was achieved is plotted. The grey ribbons represent the median, interquartile range (25th to 75th percentiles), and 95% confidence interval of 10,000 rank permutations.
Fig 4
Fig 4. Detailed analysis of the contribution of individual gene mutants to biological process predictions.
Each panel shows, for a bioprocess and either a compound (A) or a set of compounds (B-C) predicted to perturb that bioprocess, the subset of the respective chemical-genetic and L2-normalized genetic interaction profiles with signal. The importance profiles are the row-wise mean of the Hadamard product (elementwise multiplication) of each chemical-genetic interaction profile and the genetic interaction profiles for query genes with which it possessed an inner product of 2 or higher that are annotated to the GO term; they reflect the strength of each strain’s contribution to the bioprocess prediction. For all panels, a query gene from the genetic interaction network was selected if it contributed to the importance score calculation for any selected compound; query genes were ordered from left to right in ascending order of their inner products (or their average, for B-C) with the selected chemical-genetic interaction profile(s). Each strain (row) was included if it passed at least one of three criteria: 1) the magnitude of its mean genetic interaction score across the selected query genes exceeded 0.04; 2) the magnitude of its chemical-genetic interaction score (for B-C, the mean of such scores) exceeded 2.5; or 3) its importance score exceeded 0.1 (for B-C, the mean of such scores). (A) Schematic showing the prediction of the “mRNA transport” bioprocess (GO:0051028) for chemical compound NPD4142. (B) Schematic showing the prediction of “CVT pathway” (FDR < 1%) for compounds whose top prediction was to that term. (C) Schematic showing the prediction of “tubulin complex assembly” (FDR <1%).
Fig 5
Fig 5. Global visualization of the contribution of chemical-genetic interactions to CG-TARGET bioprocess predictions.
Chemical-genetic interaction profiles and their corresponding importance score profiles (see Fig 4 legend) were gathered for each of 130 diverse compounds from the high confidence set (FDR ≤ 25%) and their associated top bioprocess predictions. Importance is plotted as a function of chemical-genetic interaction score. One thousand points from the regions of lowest density (white) are plotted, with only density plotted in the remaining higher-density regions. Density increases in order of white, yellow, green, and violet. The shaded region highlights strains with strong negative (≤ –5) chemical-genetic interactions and no contribution (± 0.1) to a compound’s top bioprocess prediction.
Fig 6
Fig 6. In vivo experimental validation of cell cycle-related biological process predictions.
Phenotypic validation of cell cycle-related predictions, performed on drug-hypersensitive yeast treated with solvent control (DMSO) or compounds predicted to perturb the cell cycle. Out of 17 compounds predicted to arrest cells in G2/M phase, data are shown for the 6 that exhibited increases in the relevant phenotypes in any and all assays. Data for NPD7834 are also shown. (A) Differential interference contrast microscopy (DIC) and fluorescence upon DAPI staining showing bud size and DNA localization, respectively, after compound treatment. The scale bar represents a distance of 5 μm. (B) FACS analysis of cell populations in different cell cycle phases at 0, 2, and 4 hours after compound treatment. The green curve overlay represents the estimated cell population in G1, S and G2/M phases. (C) Budding index percentages induced by treatment with compound or solvent control.
Fig 7
Fig 7. In vitro experimental validation of “tubulin complex assembly” biological process predictions.
(A) In vitro inhibition of tubulin polymerization by compounds predicted to perturb “tubulin complex assembly” (FDR < 1%; red) compared to randomly-selected negative control compounds with high-confidence predictions to bioprocesses not related to chromosome segregation, kinetochore, spindle assembly, and microtubules (blue). Vmax values reflecting the maximum rate of tubulin polymerization for each compound from independent replicate experiments are plotted. Assay positive and negative control compounds are colored grey. (B) Structural similarity-based hierarchical clustering of compounds tested in (A). Single linkage was used in combination with (1 –structural similarity) as the distance metric; as such, the structural similarity of the two most similar compounds at each junction can be inferred directly from the dendrogram. Compounds predicted to perturb “tubulin complex assembly” (FDR < 1%) are in bold, and known microtubule-perturbing agents are marked with an asterisk. Structural similarity was calculated as the Braun-Blanquet similarity coefficient on all-shortest-path chemical fingerprints of length 8 (see Materials and Methods).

Similar articles

See all similar articles

Cited by 3 articles


    1. Giaever G, Shoemaker DD, Jones TW, Liang H, Winzeler EA, Astromoff A, et al. Genomic profiling of drug sensitivities via induced haploinsufficiency. Nat Genet. 1999. March;21(3):278–83. 10.1038/6791 - DOI - PubMed
    1. Parsons AB, Brost RL, Ding H, Li Z, Zhang C, Sheikh B, et al. Integration of chemical-genetic and genetic interaction data links bioactive compounds to cellular target pathways. Nat Biotechnol. 2004. January;22(1):62–9. 10.1038/nbt919 - DOI - PubMed
    1. Parsons AB, Lopez A, Givoni IE, Williams DE, Gray CA, Porter J, et al. Exploring the mode-of-action of bioactive compounds by chemical-genetic profiling in yeast. Cell. 2006. August 11;126(3):611–25. 10.1016/j.cell.2006.06.040 - DOI - PubMed
    1. Hillenmeyer ME, Fung E, Wildenhain J, Pierce SE, Hoon S, Lee W, et al. The chemical genomic portrait of yeast: uncovering a phenotype for all genes. Science. 2008. April 18;320(5874):362–5. 10.1126/science.1150021 - DOI - PMC - PubMed
    1. Costanzo M, Baryshnikova A, Bellay J, Kim Y, Spear ED, Sevier CS, et al. The genetic landscape of a cell. Science. 2010. January 22;327(5964):425–31. 10.1126/science.1180823 - DOI - PMC - PubMed

Publication types

MeSH terms